MeshAnything: 3D generator — what it does and what you need to run it

MeshAnything is a 3D generator that turns input data into polygonal meshes. The authors position it as a tool capable of creating meshes that resemble the work of 3D modelers, using autoregressive transformers. The project was presented at ICLR 2025.

What it does

• Generates 3D meshes from input data — the authors note that the system works on the principle of "from anything to mesh".
• Uses an autoregressive model based on a transformer architecture.
• As described by the authors, generating a single mesh takes about 30 seconds on an A6000 accelerator.

What you need to run it

• Platform: as described by the authors, the environment was tested on Ubuntu 22 with CUDA 11.8. CUDA support is required.
• VRAM: the authors state that generating a mesh requires about 7 GB of VRAM on an A6000 accelerator.
• Language: Python.
• License: non-standard. The authors do not include a typical open source license file, so you need to check the terms of use manually in the repository.
• Code: open source, available at github.com/buaacyw/MeshAnything. Last updated on April 28, 2025.

Who it is for

• Researchers in generative 3D and machine learning who want to study or refine the autoregressive approach to mesh generation.
• Developers willing to run the model on machines with CUDA-compatible accelerators and 7 GB or more of VRAM.
• Those looking for open source code for mesh generation but ready to deal with a non-standard license on their own.

The tool is not suitable for users without access to discrete NVIDIA GPUs with CUDA support, nor for those expecting a ready-made application with a graphical interface — the authors provide only Python code.

MeshAnything is an open source research project that offers one possible approach to polygonal mesh generation. Before using it, carefully review the license terms, as they differ from standard open source licenses.